Musical Genres: Beating to the Rhythms of Different Drums Debora C. Correa 1 , Jose H. Saito 2 and Luciano da F. Costa 1 1 Institute of Physics of Sao Carlos - University of Sao Paulo, Avenue Trabalhador Sao Carlense 400, Caixa Postal 369, CEP 13560-970, Sao Carlos, Sao Paulo, Brazil 2 Computer Departament - Federal University of Sao Carlos, Rodovia Washington Luis, km 235,SP-310, CEP 13565-905, Sao Carlos, Sao Paulo, Brazil 3 National Institute of Science and Technology for Complex Systems, Brazil E-mail: [email protected], luciano.if.sc.usp.br Abstract. Online music databases have increased signicantly as a consequence of the rapid growth of the Internet and digital audio, requiring the development of faster and more efficient tools for music content analysis. Musical genres are widely used to organize music collections. In this paper, the problem of automatic music genre classification is addressed by exploring rhythm-based features obtained from a respective complex network representation. A Markov model is build in order to analyse the temporal sequence of rhythmic notation events. Feature analysis is performed by using two multivariate statistical approaches: principal component analysis (unsupervised) and linear discriminant analysis (supervised). Similarly, two classifiers are applied in order to identify the category of rhythms: parametric Bayesian classifier under gaussian hypothesis (supervised), and agglomerative hierarchical clustering (unsupervised). Qualitative results obtained by Kappa coefficient and the obtained clusters corroborated the effectiveness of the proposed method. arXiv:0911.3842v1 [physics.data-an] 19 Nov 2009
36
Embed
Musical Genres: Beating to the Rhythms of Different Drums
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Musical Genres: Beating to the Rhythms of
Different Drums
Debora C. Correa1, Jose H. Saito2 and Luciano da F. Costa1
1 Institute of Physics of Sao Carlos - University of Sao Paulo, Avenue TrabalhadorSao Carlense 400, Caixa Postal 369, CEP 13560-970, Sao Carlos, Sao Paulo, Brazil2 Computer Departament - Federal University of Sao Carlos, Rodovia WashingtonLuis, km 235,SP-310, CEP 13565-905, Sao Carlos, Sao Paulo, Brazil3 National Institute of Science and Technology for Complex Systems, Brazil
Abstract. Online music databases have increased signicantly as a consequence ofthe rapid growth of the Internet and digital audio, requiring the development offaster and more efficient tools for music content analysis. Musical genres are widelyused to organize music collections. In this paper, the problem of automatic musicgenre classification is addressed by exploring rhythm-based features obtained froma respective complex network representation. A Markov model is build in orderto analyse the temporal sequence of rhythmic notation events. Feature analysisis performed by using two multivariate statistical approaches: principal componentanalysis (unsupervised) and linear discriminant analysis (supervised). Similarly, twoclassifiers are applied in order to identify the category of rhythms: parametric Bayesianclassifier under gaussian hypothesis (supervised), and agglomerative hierarchicalclustering (unsupervised). Qualitative results obtained by Kappa coefficient and theobtained clusters corroborated the effectiveness of the proposed method.
arX
iv:0
911.
3842
v1 [
phys
ics.
data
-an]
19
Nov
200
9
2
1. Introduction
Musical databases have increased in number and size continuously, paving the way to
large amounts of online music data, including discographies, biographies and lyrics.
This happened mainly as a consequence of musical publishing being absorbed by the
Internet, as well as the restoration of existing analog archives and advancements of web
technologies. As a consequence, more and more reliable and faster tools for music content
analysis, retrieval and description, are required, catering for browsing, interactive access
and music content-based queries. Even more promising, these tools, together with the
respective online music databases, have open new perspectives to basic investigations in
the field of music.
Within this context, music genres provide particularly meaningful descriptors given
that they have been extensively used for years to organize music collections. When a
musical piece becomes associated to a genre, users can retrieve what they are searching
in a much faster manner. It is interesting to notice that these new possibilities of
research in music can complement what is known about the trajectories of music genres,
their history and their dynamics [1]. In an ethnographic manner, music genres are
also particularly important because they express the general identity of the cultural
foundations in which they are comprised [2]. Music genres are part of a complex interplay
of cultures, artists and market strategies to define associations between musicians and
their works, making the organization of music collections easier [3]. Therefore, musical
genres are of great interest because they can summarise some shared characteristics
in music pieces. As indicated by [4], music genre is probably the most common
description of music content, and its classification represents an appealing topic in Music
Information Retrieval (MIR) research.
Despite their ample use, music genres are not a clearly defined concept, and their
boundaries remain fuzzy [3]. As a consequence, the development of such taxonomy is
controversial and redundant, representing a challenging problem. Pachet and Cazaly [5]
demonstrated that there is no general agreement on musical genre taxonomies, which
can depend on cultural references. Even widely used terms such as rock, jazz, blues and
pop are not clear and firmly defined. According to [3], it is necessary to keep in mind
what kind of music item is being analysed in genre classification: a song, an album, or
an artist. While the most natural choice would be a song, it is sometimes questionable
to classify one song into only one genre. Depending on the characteristics, a song can be
classified into various genres. This happens more intensively with albums and artists,
since nowadays the albums contain heterogeneous material and the majority of artists
tend to cover an ample range of genres during their careers. Therefore, it is difficult
to associate an album or an artist with a specific genre. Pachet and Cazaly [5] also
mention that the semantic confusion existing in the taxonomies can cause redundancies
that probably will not be confused by human users, but may hardly be dealt with
by automatic systems, so that automatic analysis of the musical databases becomes
essential. However, all these critical issues emphasize that the problem of automatic
3
classification of musical genres is a nontrivial task. As a result, only local conclusions
about genre taxonomy are considered [5].
As other problems involving pattern recognition, the process of automatic
classification of musical genres can usually be divided into the following three main steps:
representation, feature extraction and the classifier design [6, 7]. Music information can
be described by symbolic representation or based on acoustic signals [8]. The former
is a high-level kind of representation through music scores, such as MIDI, where each
note is described in terms of pitch, duration, start time and end time, and strength.
Acoustic signals representation is obtained by sampling the sound waveform. Once the
audio signals are represented in the computer, the objective becomes to extract relevant
features in order to improve the classification accuracy. In the case of music, features
may belong to the main dimensions of it including melody, timbre, rhythm and harmony.
After extracting significant features, any classification scheme may be used. There
are many previous works concerning automatic genre classification in the literature
[9, 10, 11, 12, 13, 14, 15, 8, 16, 17, 18, 19].
An innovative approach to automatic genre classification is proposed in the current
work, in which musical features are referred to the temporal aspects of the songs: the
rhythm. Thus, we propose to identify the genres in terms of their rhythmic patterns.
While there is no clear definition of rhythm [3], it is possible to relate it with the idea
of temporal regularity. More generally speaking, rhythm can be simply understood as
a specific pattern produced by notes differing in duration, pause and stress. Hence, it
is simpler to obtain and manipulate rhythm than the whole melodic content. However,
despite its simplicity, the rhythm is genuine and intuitively characteristic and intrinsic
to musical genres, since, for example, it can be use to distinguish between rock music
and rhythmically more complex music, such as salsa. In addition, the rhythm is largely
independent on the instrumentation and interpretation.
A few related works that use rhythm as features in automatic genre recognition
can be found in the literature. The work of Akhtaruzzaman [20], in which rhythm
is analysed in terms of mathematical and geometrical properties, and then fed to a
system for classification of rhythms from different regions. Karydis [21] proposed to
classify intra-classical genres with note pitch and duration features, obtained from their
histograms. In [22, 23], a review of existing automatic rhythm description systems are
presented. The authors say that despite the consensus on some rhtyhm concepts, there is
not a single representation of rhythm that would be applicable for different applications,
such as tempo and meter induction, beat tracking, quantization of rhythm and so on.
They also analysed the relevance of these descriptors by mesuring their performance in
genre classification experiments. It has been observed that many of these approaches
lack comprehensiveness because of the relatively limited rhythm representations which
have been adopted [3].
In the current study, an objective and systematic analysis of rhythm is provided.
The main motivation is to study similar and different characteristics of rhythms in
terms of the occurrence of sequences of events obtained from rhythmic notations. First,
4
the rhythm is extracted from MIDI databases and represented as graphs or networks
[24, 25]. More specifically, each type of note (regarding their duration) is represented
as a node, while the sequence of notes define the links between the nodes. Matrices
of probability transition are extracted from these graphs and used to build a Markov
model of the respective musical piece. Since they are capable of systematically modeling
the dynamics and dependencies between elements and subelements [26], Markov models
are frequently used in temporal pattern recognition applications, such as handwriting,
speech, and music [27]. Supervised and an unsupervised approaches are then applied
which receive as input the properties of the transition matrices and produce as output
the most likely genre. Supervised classification is performed with the Bayesian classifier.
For the unsupervised approach, a taxonomy of rhythms is obtained through hierarchical
clustering. The described methodology is applied to four genres: blues, bossa nova,
reggae and rock, which are well-known genres representing different tendencies. A series
of interesting findings are reported, including the ability of the proposed framework
to correctly identify the musical genres of specific musical pieces from the respective
rhythmic information.
This paper is organized as follows: section 2 describes the methodology, including
the classification methods; section 3 presents the obtained results as well as their
discussion, and section 4 contains the concluding remarks and future works.
2. Materials and Methods
Some basic conceps about complex networks as well as the proposed methodology are
presented in this section.
2.1. Systems Representation by Complex Networks
A complex network is a graph exhibiting intricate structure when compared to regular
and uniformly random structures. There are four main types of complex networks:
weighted and unweighted digraphs and weighted and unweighted graphs. The operations
of simmetry and thresholding can be used to transform a digraph into a graph and
a weighted graph (or weighted digraph) into an unweighted one, respectively [28]. A
weighted digraph (or weighted direct graph) G can be defined by the following elements:
- Vertices (or nodes). Each vertex is represented by an integer number i = 1, 2, ..., N ;
N(G) is the vertex set of digraph G and N indicates the total number of vertices
(|N(G)|).- Edges (or links). Each edge has the form (i, j) indicating a connection from vertex
i to vertex j. The edge set of digraph G is represented by ε(G), and M is the total
number of edges.
- The mapping ω : ε(G) 7→ R, where R is the set of weight values. Each edge
(i, j) has a weight ω(i, j) associated to it. This mapping does not exist in unweighted
digraphs.
5
Table 1. Graphs and Digraphs Basic ConceptsGraphs Digraphs
Adjacency Two vertices i e j are Concepts ofadjacent or neighbors if aij = 1 predecessors and successors. If
aij 6= 0, i is predecessor ofj and j is successor of i.Predecessors e successors
as adjacent vertices.Neighborhood Represented by v(i), meaning Also represented by v(i).
the set of vertices that areneighbors to vertex i.
Vertex degree Represented by ki, gives There are two kinds of degrees: in-degree kini
the number of connected edges indicating the number of incoming edges;to vertex i. It is computed as: and out-degree kout
i
ki =∑j
aij =∑j
aji indicating the number of outgoing edges:
kini =
∑j
aji kouti =
∑j
aij
The total degree is defined as ki = kini + kout
i
Average degree Average of ki considering all It is the same for in- and out- degrees.network vertices.
〈k〉 = 1N
∑i
ki = 1N
∑ij
aij 〈kout〉 =⟨kin⟩
= 1N
∑ij
aij
Undirected graphs (weighted or unweighted) are characterized by the fact that
their edges have not orientation. Therefore, an edge (i, j) in such a graph necessarily
implies a connection from vertex i to vertex j and from vertex j to vertex i. A weighted
digraph can be represented in terms of its weight matrices W . Each element of W , wji,
associates a weight to the connection from vertex i to vertex j. The table 1 summarizes
some fundamental concepts about graphs and digraphs [28].
For weighted networks, a quantity called strength of vertex i is used to express
the total sum of weights associated to each node. More specifically, it corresponds to
the sum of the weights of the respective incoming edges (sini =
∑jwji) (in-strength) or
outgoing edges (souti =
∑jwij) (out-strength) of vertex i.
Another interesting measurement of local connectivity is the clustering coefficient.
This feature reflects the cyclic structure of networks, i.e. if they have a tendency to form
sets of densely connected vertices. For digraphs, one way to calculate the clustering
coefficient is: let mi be the number of neighbors of vertex i and li be the number of
connections between the neighbors of vertex i; the clustering coefficient is obtained as
cc(i) = li/mi(mi − 1).
2.2. Data Description
In this work, four music genres were selected: blues, bossa nova, reggae and rock.
These genres are well-known and represent distinct major tendencies. Music samples
belonging to these genres are available in many collections in the Internet, so it was
6
possible to select one hundred samples to represent each one of them. These samples
were downloaded in MIDI format. This event-like format contains instructions (such as
notes, instruments, timbres, rhythms, among others) which are used by a synthesizer
during the creation of new musical events [29]. The MIDI format can be considered a
digital musical score in which the instruments are separated into voices.
In order to edit and analyse the MIDI scores, we applied the software for music
composition and notation called Sibelius (http://www.sibelius.com). For each
sample, the voice related to the percussion was extracted. The percussion is inherently
suitable to express the rhythm of a piece. Once the rhythm is extracted, it becomes
possible to analyse all the involved elements. The MIDI Toolbox for Matlab was used
[30]. This Toolbox is free and contains functions to analyse and visualize MIDI files
in the Matlab computing environment. When a MIDI file is read with this toolbox,
a matrix representation of note events is created. The columns in this matrix refer to
many types of information, such as: onset (in beats), duration (in beats), MIDI channel,
MIDI pitch, velocity, onset (in seconds) and duration (in seconds). The rows refer to
the individual note events, that is, each note is described in terms of its duration, pitch,
and so on.
Only the note duration (in beats) has been used in the current work. In fact, the
durations of the notes, respecting the sequence in which they occur in the sample, are
used to create a digraph. Each vertex of this digraph represents one possible rhythm
notation, such as, quarter note, half note, eighth note, and so on. The edges reflect the
subsequent pairs of notes. For example, if there is an edge from vertex i, represented by
a quarter note, to a vertex j, represented by an eighth note, this means that a quarter
note was followed by an eighth note at least once. The thicker the edges, the larger
is the strength between these two nodes. Examples of these digraphs are showed in
Figure 1. Figure 1(a) depicts a blues sample represented by the music How blue can you
get by BB King. A bossa nova sample, namely the music Fotografia by Tom Jobim, is
illustrated in 1(b). Figure 1(c) illustrates a reggae sample, represented by the music Is
this Love by Bob Marley. Finally, Figure 1(d) shows a rock sample, corresponding the
music From Me To You by The Beatles.
2.3. Feature Extraction
Extracting features is the first step of most pattern recognition systems. Each pattern
is represented by its d features or attributes in terms of a vector in a d - dimensional
space. In a discrimination problem, the goal is to choose features that allow the pattern
vectors belonging to different classes to occupy compact and distinct regions in the
feature space, maximizing class separability. After extracting significant features, any
classification scheme may be used. In the case of music, features may belong to the
main dimensions of it including melody, timbre, rhythm and harmony.
Therefore, one of the main features of this work is to extract features from the
digraphs and use them to analyse the complexities of the rhythms, as well as to perform
Figure 1. Digraph examples of four music samples: (a) How Blue Can You Get byBB King. (b) Fotografia by Tom Jobim. (c) Is This Love by Bob Marley. (d) FromMe To You by The Beatles.
classification tasks. For each sample, a digraph is created as described in the previous
section. All digraphs have 18 nodes, corresponding to the quantity of rhythm notation
possibilities concerning all the samples, after excluding those that hardly ever happens.
This exclusion was important in order to provide an appropriate visual analysis and to
better fit the features. In fact, avoiding features that do not significantly contribute to
the analysis reduces data dimension, improves the classification performance through
a more stable representation and removes redundant or irrelevant information (in this
case, minimizes the occurrence of null values in the data matrix).
8
The features are associated with the weight matrix W . As commented in section
2.1, each element in W , wij, indicates the weight of the connection from vertex j to i, or,
in other words, they are meant to represent how often the rhythm notations follow one
another in the sample. The weight matrix W has 18 rows and 18 columns. The matrix
W is reshaped by a 1 x 324 feature vector. This is done for each one of the genre samples.
However, it was observed that some samples even belonging to different genres generated
exactly the same weight matrix. These samples were excluded. Thereby, the feature
matrix has 280 rows (all non-excluded samples), and 324 columns (the attributes).
An overview of the proposed methodology is illustrated in Figure 2. After extracting
the features, a standardization transformation is done to guarantee that the new feature
set has zero mean and unit standard deviation. This procedure can significantly improve
the resulting classification. Once the normalized features are available, the structure of
the extracted rhythms can be analysed by using two different approaches for features
analysis: PCA and LDA. We also compare two types of classification methods: Bayesian
classifier (supervised) and hierarchical clustering (unsupervised). PCA and LDA are
described in section 2.4 and the classification methods are described in section 2.5.
2.4. Feature Analysis and Redundancy Removing
Two techniques are widely used for feature analysis [31, 7, 32]: PCA (Principal
Components Analysis) and LDA (Linear discriminant Analysis). Basically, these
approaches apply geometric transformations (rotations) to the feature space with the
purpose of generating new features based on linear combinations of the original ones,
aiming at dimensionality reduction (in the case of PCA) or to seek a projection that
best separates the data (in the case of LDA). Figure3 illustrates the basic principles
underlying PCA and LDA. The direction x′ (obtained with PCA) is the best one to
represent the two classes with maximum overall dispersion. However, tt can be observed
that the densities projected along direction x′ overlap one another, making these two
classes inseparable. Differently, if direction y′, obtained with LDA, is chosen, the classes
can be easily separated. Therefore, it is said that the directions for representation are
not always also the best choice for classification, reflecting the different objectives of
PCA and LDA. [33]. Appendix A.1 and Appendix A.2 give more details about these
two techniques.
2.5. Classification Methodology
Basically, to classify means to assign objects to classes or categories according to the
properties they present. In this context, the objects are represented by attributes, or
feature vectors. There are three main types of pattern classification tasks: imposed
criteria, supervised classification and unsupervised classification. Imposed criteria is
the easiest situation in classification, once the classification criteria is clearly defined,
generally by a specific practical problem. If the classes are known in advance, the
classification is said to be supervised or by example, since usually examples (the training
9
Figure 2. Block diagram of the proposed methodology.
Figure 3. An illustration of PCA (optimizing representation) and LDA projections(optimizing classification), adapted from [33].
10
set) are available for each class. Generally, supervised classification involves two stages:
learning, in which the features are tested in the training set; and application, when
new entities are presented to the trained system. There many approaches involving
supervised classification. The current study applied the Bayesian classifier through
discriminant functions (Appendix B) in order to perform supervised classification.
The Bayesian classifier is based on the Bayesian Decision Theory and combines class
conditional probability densities (likelihood) and prior probabilities (prior knowledge)
to perform classification by assigning each object to the class with the maximum a
posteriori probability.
In unsupervised classification, the classes are not known in advance, and there is
not a training set. This type of classification is usually called clustering, in which the
objects are agglomerated according to some similarity criterion. The basic principle is
to form classes or clusters so that the similarity between the objects in each class is
maximized and the similarity between objects in different classes minimized. There are
two types of clustering: partitional (also called non-hierarchical) and hierarchical. In the
former, a fixed number of clusters is obtained as a single partition of the feature space.
Hierarchical clustering procedures are more commonly used because they are usually
simpler [7, 6, 32, 34]. The main difference is that instead of one definite partition, a
series of partitions are taken, which is done progressively. If the hierarchical clustering
is agglomerative (also known as bottom-up), the procedure starts with N objects as
N clusters and then successively merges the clusters until all the objects are joined
into a single cluster (please refer to Appendix C for more details about agglomerative
hierarchical clustering). Divisive hierarchical clustering (top-down) starts with all the
objects as one single cluster, and splits it into progressively finer subclusters.
2.5.1. Performance Measures for Classification To objectively evaluate the perfor-
mance of the supervised classification it is necessary to use quantitative criteria. Most
used criteria are the estimated classification error and the obtained accuracy. Because
of its good statistical properties, as, for example, be asymptotically normal with well
defined expressions to estimate its variances, this study also adopted the Cohen Kappa
Coefficient [35] as a quantitative measure to analyse the performance of the proposed
method. Besides, the kappa coefficient can be directly obtained from the confusion
matrix [36] (Appendix D), easily computed in supervised classification problems. The
confusion matrix is defined as:
C =
c11 c12 . . . c1c
c21. . .
......
. . .
cc1 . . . ccc
(1)
where each element cij represents the number of objects from class i classified
as class j. Therefore, the elements in the diagonal indicate the number of correct
classifications.
11
Table 2. Blues and Bossa nova art worksBlues art works
1 Albert Collins- A Good Fooll2 Albert Collins- Fake ID3 Albert King- Stormy Monday4 Andrews Sister- Boogie Woogie Bugle Boy5 B B King- Dont Answer The Door6 B B King- Get Off My Back7 B B King- Good Time8 B B King- How Blue Can You Get9 B B King- Sweet Sixteen10 B B King- The Thrill Is Gone11 B B King- Woke Up This Morning12 Barbra Streisand- Am I Blue13 Billie Piper- Something Deep Inside14 Blues In F For Monday15 Bo Diddley- Bo Diddley16 Boy Williamson- Dont Start Me Talking17 Boy Williamson- Help Me18 Boy Williamson- Keep It To Yourself19 Buddy Guy- Midnight Train20 Charlie Parker- Billies Bounce21 Count Basie- Count On The Blues22 Cream- Crossroads Blues23 Delmore Brothers- Blues Stay Away From Me24 Elmore James- Dust My Broom25 Elves Presley- A Mess Of Blues26 Etta James- At Last27 Feeling The Blues28 Freddie King- Help Day29 Freddie King- Hide Away30 Gary Moore- A Cold Day In Hell31 George Thorogood- Bad To The Bone32 Howlin Wolf- Little Red Rooster33 Janis Joplin- Piece Of My Heart34 Jimmie Cox- Before You Accuse Me35 Jimmy Smith- Chicken Shack36 John Lee Hooker- Boom Boom Boom37 John Lee Hooker- Dimples38 John Lee Hooker- One Bourbon One Scotch One Beer39 Johnny Winter- Good Morning Little School Girl40 Koko Taylor- Hey Bartender41 Little Walter- Juke42 Louis Jordan- Let Good Times Roll43 Miles Davis- All Blues44 Ray Charles- Born To The Blues45 Ray Charles- Crying Times46 Ray Charles- Georgia On My Mind47 Ray Charles- Hit The Road Jack48 Ray Charles- Unchain My Heart49 Robert Johnson- Dust My Broom50 Stevie Ray Vaughan- Cold Shot51 Stevie Ray Vaughan- Couldnt Stand The Weather52 Stevie Ray Vaughan- Dirty Pool53 Stevie Ray Vaughan- Hillbillies From Outer Space54 Stevie Ray Vaughan- I Am Crying55 Stevie Ray Vaughan- Lenny56 Stevie Ray Vaughan- Little Wing57 Stevie Ray Vaughan- Looking Out The Window58 Stevie Ray Vaughan- Love Struck Baby59 Stevie Ray Vaughan- Manic Depression60 Stevie Ray Vaughan- Scuttle Buttin61 Stevie Ray Vaughan- Superstition62 Stevie Ray Vaughan- Tell Me63 Stevie Ray Vaughan- Voodoo Chile64 Stevie Ray Vaughan- Wall Of Denial65 T Bone Walker- Call It Stormy Monday66 The Blues Brothers- Everybody Needs Somebody To Love67 The Blues Brothers- Green Onions68 The Blues Brothers- Peter Gunn Theme69 The Blues Brothers- Soulman70 W C Handy- Memphis Blues
Bossa nova art works
71 Antonio Adolfo- Sa Marina72 Barquinho73 Caetano Veloso- Menino Do Rio74 Caetano Veloso- Sampa75 Celso Fonseca- Ela E Carioca76 Celso Fonseca- Slow Motion bossa nova77 Chico Buarque Els Soares- Facamos78 Chico Buarque Francis Hime- Meu Caro Amigo79 Chico Buarque Quarteto Em Si- Roda Viva80 Chico Buarque- As Vitrines81 Chico Buarque- Construcao82 Chico Buarque- Desalento83 Chico Buarque- Homenagem Ao Malandro84 Chico Buarque- Mulheres De Athenas85 Chico Buarque- Ole O La86 Chico Buarque- Sem Fantasia87 Chico Buarque- Vai Levando88 Dick Farney- Copacaba89 Elis Regina- Alo Alo Marciano90 Elis Regina- Como Nossos Pais91 Elis Regina- Na Batucada Da Vida92 Elis Regina- O Bebado E O Equilibrista93 Elis Regina- Romaria94 Elis Regina- Velho Arvoredo95 Emilio Santiago- Essa Fase Do Amor96 Emilio Santiago- Esta Tarde Vi Voller97 Emilio Santiago- Saigon98 Emilio Santiago- Ta Tudo Errado99 Gal Costa- Canta Brasil100 Gal Costa- Para Machucar Meu Coracao101 Gal Costa- Pra Voce102 Gal Costa- Um Dia De Domingo103 Jair Rodrigues- Disparada104 Joao Bosco- Corsario105 Joao Bosco- De Frente Para O Crime106 Joao Bosco- Jade107 Joao Bosco- Risco De Giz108 Joao Gilberto- Corcovado109 Joao Gilberto- Da Cor Do Pecado110 Joao Gilberto- Um Abraco No Bonfa111 Luiz Bonfa- De Cigarro Em Cigarro112 Luiz Bonfa- Manha De Carnaval113 Marcos Valle- Preciso Aprender A Viver So114 Marisa Monte- Ainda Lembro115 Marisa Monte- Amor I Love You116 Marisa Monte- Ando Meio Desligado117 Tom Jobim- Aguas De Marco118 Tom Jobim- Amor E Paz119 Tom Jobim- Brigas Nunca Mais120 Tom Jobim- Desafinado121 Tom Jobim- Fotografia122 Tom Jobim- Garota De Ipanema123 Tom Jobim- Meditacao124 Tom Jobim- Samba Do Aviao125 Tom Jobim- Se Todos Fossem Iguais A Voce126 Tom Jobim- So Tinha De Ser Com Voce127 Tom Jobim- Vivo Sonhando128 Tom Jobim- Voce Abusou129 Tom Jobim- Wave130 Toquinho- Agua Negra Da Lagoa131 Toquinho- Ao Que Vai132 Toquinho- Este Seu Olhar133 Vinicius De Moraes- Apelo134 Vinicius De Moraes- Carta Ao Tom135 Vinicius De Moraes- Minha Namorada136 Vinicius De Moraes- O Morro Nao Tem Vez137 Vinicius De Moraes- Onde Anda Voce138 Vinicius De Moraes- Pela Luz Dos Olhos Teus139 Vinicius De Moraes- Samba Em Preludio140 Vinicius De Moraes-Tereza Da Praia
3. Results and Discussion
As mentioned before, four musical genres were used in this study: blues, bossa nova,
reggae and rock. It was selected music art works from diverse artists, as presented in
Tables 2 and 3. Different colors were chosen to represent the genres (color red for genre
blues, green for bossa nova, cyan for reggae and pink for rock), in order to provide a
better visualization and discussion of the results.
Once we want to reduce data dimensionality, it is necessary to set the suitable
12
Table 3. Reggae and Rock art worksReggae art works
141 Ace Of Bass- All That She Wants142 Ace Of Bass- Dont Turn Around143 Ace Of Bass- Happy Nation144 Armandinho- Pela Cor Do Teu Olho145 Armandinho- Sentimento146 Big Mountain- Baby I Love Your Way147 Bit Meclean- Be Happy148 Bob Marley- Africa Unite149 Bob Marley- Buffalo Soldier150 Bob Marley- Exodus151 Bob Marley- Get Up Stand Up152 Bob Marley- I Shot The Sheriff153 Bob Marley- Iron Lion Zion154 Bob Marley- Is This Love155 Bob Marley- Jammin156 Bob Marley- No Woman No Cry157 Bob Marley- Punky Reggae Party158 Bob Marley- Root Rock159 Bob Marley- Satisfy160 Bob Marley- Stir It Up161 Bob Marley- Three Little Birds162 Bob Marley- Waiting In The Van163 Bob Marley- War164 Bob Marley- Wear My165 Bob Marley- Zimbabwe166 Cidade Negra- A Cor Do Sol167 Cidade Negra- A Flexa E O Vulcao168 Cidade Negra- A Sombra Da Maldade169 Cidade Negra- Aonde Voce Mora170 Cidade Negra- Eu Fui Eu Fui171 Cidade Negra- Eu Tambem Quero Beijar172 Cidade Negra- Firmamento173 Cidade Negra- Girassol174 Cidade Negra- Ja Foi175 Cidade Negra- Mucama176 Cidade Negra- O Ere177 Cidade Negra- Pensamento178 Dazaranhas- Confesso179 Dont Worry180 Flor Do Reggae181 Inner Circle- Bad Boys182 Inner Circle- Sweat183 Jimmy Clif- I Can See Clearly Now184 Jimmy Clif- Many Rivers To Cross185 Jimmy Clif- Reggae Night186 Keep On Moving187 Manu Chao- Me Gustas Tu188 Mascavo- Anjo Do Ceu189 Mascavo- Asas190 Natiruts- Liberdade Pra Dentro Da Cabeca191 Natiruts- Presente De Um Beija Flor192 Natiruts- Reggae Power193 Nazarite Skank194 Peter Tosh- Johnny B Goode195 Shaggy- Angel196 Shaggy- Bombastic197 Shaggy- It Wasnt Me198 Shaggy- Strenght Of A Woman199 Sublime- Badfish200 Sublime- D Js201 Sublime- Santeria202 Sublime- Wrong Way203 Third World- Now That We Found Love204 Tribo De Jah- Babilonia Em Chamas205 Tribo De Jah- Regueiros Guerreiros206 Tribo De Jah- Um So Amor207 U B40- Bring Me Your Cup208 U B40- Home Girl209 U B40- To Love Somebody210 Ub40- Red Red Wine
Rock art works
211 Aerosmith- Kings And Queens212 Aha- Stay On These Roads213 Aha- Take On Me214 Aha- Theres Never A Forever Thing215 Beatles- Cant Buy Me Love216 Beatles- From Me To You217 Beatles- I Want To Hold Your Hand218 Beatles- She Loves You219 Billy Idol- Dancing With Myself220 Cat Steves- Another Saturday Night221 Creed- My Own Prison222 Deep Purple- Deamon Eye223 Deep Purple- Hallelujah224 Deep Purple- Hush225 Dire Straits- Sultan Of Swing226 Dire Straits- Walk Of Life227 Duran Duran- A View To A Kill228 Eric Clapton- Cocaine229 Europe- Carrie230 Fleetwood Mac- Dont Stop231 Fleetwood Mac- Dreams232 Fleetwood Mac- Gold Dust Woman233 Foo Fighters- Big Me234 Foo Fighters- Break Out235 Foo Fighters- Walking After You236 Men At Work- Down Under237 Men At Work- Who Can It Be Now238 Metallica- Battery239 Metallica- Fuel240 Metallica- Hero Of The Day241 Metallica- Master Of Puppets242 Metallica- My Friendof Misery243 Metallica- No Leaf Clover244 Metallica- One245 Metallica- Sad But True246 Pearl Jam- Alive247 Pearl Jam- Black248 Pearl Jam- Jeremy249 Pet Shop Boys- Go West250 Pet Shop Boys- One In A Million251 Pink Floyd- Astronomy Domine252 Pink Floyd- Have A Cigar253 Pink Floyd- Hey You254 Queen- Another One Bites The Dust255 Queen- Dont Stop Me Now256 Queen- I Want It All257 Queen- Play The Game258 Queen- Radio Gaga259 Queen- Under Pressure260 Red Hot Chilli Pepers- Higher Ground261 Red Hot Chilli Pepers- Otherside262 Red Hot Chilli Pepers- Under The Bridge263 Rolling Stones- Angie264 Rolling Stones- As Tears Go By265 Rolling Stones- Satisfaction266 Rolling Stones- Street Of Love267 Steppen Wolf- Magic Carpet Ride268 Steve Winwood- Valerie269 Steve Winwood- While You See A Chance270 Tears For Fears- Shout271 The Doors- Hello I Love You272 The Doors- Light My Fire273 The Doors- Love Her Madly274 U2- Elevation275 U2- Ever Lasting Love276 U2- When Love Comes To Town277 Van Halen- Dance The Night Away278 Van Halen- Dancing With The Devil279 Van Halen- Jump280 Van Halen- Panama
number of principal components that will represent the new features. Not surprisingly,
in a high dimensional space the classes can be easily separated. On the other hand, high
dimensionality increases complexity, making the analysis of both extracted features and
classification results a difficult task. One approach to get the ideal number of principal
components is to verify how much of the data variance is preserved. In order to do
so, the l first eigenvalues (l is the quantity of principal components to be verified) are
summed up and the result is divided by the the sum of all the eigenvalues. If the
result of this calculation is a value equal or greater than 0.75, it is said that these
13
number of components (or new features) preserves at least 75% of the data variance,
which is often enough for classification purposes. When PCA was applied to the
normalized rythm features, as illustrated in Figure 4, it was observed that 20 principal
components preserved 76% of the variance of the data. That is, it is possible to reduce
the data dimensionality from 364-D to 20-D without a significant loss of information.
Nevertheless, as will be shown in the following, depending on the classifier and how the
classification task was performed, different number of components were required in each
situation in order to achieve suitable results. Despite the fact of preserving only 32% of
the variance, Figure 4 shows the first three principal components, that is, the first three
new features obtained with PCA. Figure 4(a) shows the first and second features and
Figure 4(b) shows the first and third features. It is noted that the classes are completely
overlapped, making the problem of automatic classification a nontrivial task.
14
(a)
(b)
Figure 4. The new first three feaures obtained by PCA. (a) First and second axes.(b) First and third axes.
15
Table 4. PCA kappa and accuracy for the Quadratic Bayesian ClassifierKappa Variance Accuracy Performance
Table 9. Confusion Matrix for the Quadratic Bayesian Classifier using LDA andRe-Substitution
Blues Bossa nova Reggae Rock
Blues 58 6 4 2Bossa nova 3 52 7 8Reggae 8 8 44 10Rock 4 7 5 54
Table 10. Misclassified art works using LDA with Quadratic Bayesian Classifier andRe-Substitution
Blues as Bossa nova 2 30 52 54 58 63Blues as Reggae 5 17 28 69Blues as Rock 6 51Bossa nova as Blues 79 80 116Bossa nova as Reggae 90 96 97 105 112 122 131Bossa Bova as Rock 76 78 83 98 100 103 107 109Reggae as Blues 144 152 166 170 188 190 193 203Reggae as Bossa nova 158 159 162 164 169 181 197 208Reggae as Rock 146 156 157 160 168 175 180 189 198 210Rock as Blues 238 239 245 278Rock as Bossa nova 233 237 254 262 264 266 276Rock as Reggae 219 228 255 261 263
Despite the value of kappa and its variance being the same using LDA with re-
substitution and PCA with re-substitution, the two confusion matrix are strongly
distinct each other. In the first case, demostrated in Table 9, the misclassified art
works are well distributed among the four classes, while with PCA (Table 5) they
are concentrated in one class, represented by the reggae genre. The results obtained
with LDA technique are particularly promising because they reflect the nature of the
data. Although widely used, terms such as rock, reggae or pop often remain loosely
defined [3]. Yet, it is worthwhile to remember that the intensity of the beat, which is
a very important aspect of the rhythm has not been considered in this work. This
means that analysing rhythm only through notations, as currently proposed, could
poise difficulties even for human experts. These misclassified art works have similar
properties described in terms of rhythm notations and, as a result, they generate
similar weight matrices. Therefore, the proposed methodology, although requiring some
complementations, seems to be a significant contribution towards the development of
viable alternative approach to automatic genre classification.
The misclassified art works of the confusion matrix in Table 9 are identified in Table
10.
The results for the Linear Bayesian Classifier using LDA are shown in Table 11.
20
Table 11. LDA kappa and accuracy for the Linear Bayesian ClassifierKappa Variance Accuracy Performance
In fact, they are closely similar to those obtained by the Quadratic Bayesian Classifier
(Table 8).
As mentioned in section Appendix A.2, linear discriminant analysis also allows us
to quantify the intra and interclass dispersion of the feature matrix through functionals
such as the trace and determinant computed from the scatter matrices [37]. The overall
intraclass scatter matrix, hence Sintra; the intraclass scatter matrix for each class, hence
SintraBlues, SintraBossaNova, SintraReggae and SintraRock; the interclass scatter matrix, hence,
Sinter; and the overall separability index, hence(S−1
intra ∗ Sinter
), were computed. Their
respective traces are:
trace(Sintra) = 499.526
trace(SintraBlues) = 138.615
trace(SintraBossaNova) = 119.302
trace(SintraReggae) = 98.327
trace(SintraRock) = 143.280
trace(Sinter) = 21.598
trace(S−1
intra ∗ Sinter
)= 3.779
Two important observations are worth mentioning. First, these traces emphasise
the difficulty of this classification problem: the traces of the intraclass scatter matrices
are too high, and the trace of the interclass scatter matrix together with the overal
separability index, too small. This confirms that the four classes are overlapping
completely. Second, the smaller intraclass trace is related to the reggae genre (it is
the most compact class). This may justify why in the experiments art works belonging
to reggae were more frequently 90%-100% correctly classified.
The PCA and LDA approaches help to identify which features contribute the most
to the classification. This is an interesting analysis that can be performed by verifying
the strength of each element in the first eigenvectors, and then associating those elements
with the original features. Within the current study, it was figured out that the first ten
sequences of rhythm notations that most contributed to separation correspond to those
illustrated in Figure 7. In the case of the first and second eigenvectors obtained by PCA
and LDA, the ten elements with higher values were selected, and the indices of these
elements were associated with the sequences in the original weight matrix. Figure 7(a)
and 7(b) shows the resulting sequences according to the first and second eigenvectors
of PCA. The thickness of the edges is set by the value of the corresponding element in
the eigenvector. It is interesting that these sequences are those that mostly frequently
happen in the rhythms from all four genres studied here. That is, they correspond
21
to the elements that play the greatest role in representing the rhythms. Therefore, it
can be said that these are the ten most representative sequences, when the first and
second eigenvectors of PCA are taken into account. Triples of eighth and sixteenth
notes are particularly important in blues and reggae genres. Similarly, Figure 7(c) and
Figure 7(d) show the resulting sequences according to the first and second eigenvectors
of LDA. Differently from those obtained by PCA, these sequences are not common
to all the rhythms, but they must happen with distinct frequency within each genre.
Thus, they are referred here as the ten most discriminative sequences, when the first
and second eigenvectors of LDA are taken into account.
22
(a) (b)
(c) (d)
Figure 7. Ten most significant sequences of rhythm notations according to: (a) PCAfirst eigenvector (b) PCA second eigenvector (c) LDA first eigenvector (d) LDA secondeigenvector
23
Table 12. Confusion Matrix for the Agglomerative Hierarchical ClusteringClasse 1 Classe 2 Classe 3 Classe 4
Figure 12. The detailed dendrogram for cluster in cyan. The objects were specifiedhere, from left to right: 6 231 160 157 189 200 107 217 274 168 275 142 153187 51 236 267 83 100 143 249 257 103 146 273 180 210 102 176 17 97 109228 198 255 196 261 192 219 56 202 150 139 76 218 229 and 194.
4. Concluding Remarks
Automatic music genre classification has become a fundamental topic in music research
since genres have been widely used to organize and describe music collections. They also
reveal general identities of different cultures. However, music genres are not a clearly
defined concept so that the development of a non-controversial taxonomy represents a
challenging, non-trivial taks.
Generally speaking, music genres summarize common characteristics of musical
pieces. This is particular interesting when it is used as a resource for automatic
classification of pieces. In the current paper, we explored genre classification while taking
into account the musical temporal aspects, namely the rhythm. We considered pieces
of four musical genres (blues, bossa nova, reggae and rock), which were extracted from
MIDI files and modeled as networks. Each node corresponded to one rhythmc notation,
and the links were defined by the sequence in which they occurred along time. The idea
of using static nodes (nodes with fixed positions) is particularly interesting because it
provides a primary visual identification of the differences and similarities between the
rythms from the four genres. A Markov model was build from the networks, and the
dynamics and dependencies of the rhythmic notations were estimated, comprising the
feature matrix of the data. Two different approaches for features analysis were used
(PCA and LDA), as well as two types of classification methods (Bayesian classifier and
hierarchical clustering).
Using only the first two principal componentes, the different types of rhythms were
not separable, although for the first and third axes we could observe some separation
27
between three of the classes (blues, bossa nova and reggae), while only the samples of
rock overlapped the other classes. However, taking into account that twenty components
were necessary to preserve 76% of the data variance, it would be expected that only two
or three dimensions would not be sufficient to allow suitable saparability. Notably, the
dimensionality of the problem is high, that is, the rhythms are very complex and many
dimensions (features) are necessary to separate them. This is one of the main findings
of the current work. With the help of LDA analysis, another finding was reached which
supported the assumption that the problem of automatic rhythm classification is no
trivial task. The projections obtained by considering the first and second, and first and
third axes implied in better discrimination between the four classes than that obtained
by the PCA.
Unlike PCA and LDA, agglomerative hierarchical clustering works on the original
dimensions of the data. The application of the methodology led to a substantially
better discrimination, which provides a strong evidence of the complexity of the problem
studied here. The results are promising in the sense that in each cluster is dominated
by a different genre, showing the viability of the proposed approach.
It is clear from our study that musical genres are very complex and present
redundancies. Sometimes it is difficult even for an expert to distinguish them. This
difficulty becomes more critical when only the rhythm is taken into account.
Several are the possibilities for future research implied by the reported investigation.
First, it would be interesting to use more measurements extracted from rhythm,
especially the intensity of the beats, as well as the distribution of instruments, which
is poised to improve the classification results. Another promising venue for further
investigation regards the use of other classifiers, as well as the combination of results
obtained from ensemble of distinct classifiers. In addition, it would be promising to apply
multi-labeled classification, a growing field of research in which non-disjointed samples
can be associated to one or more labels [38]. Nowadays, multi-labeled classification
methods have been increasingly required by applications such as text categorization [39],
scene classification [40], protein classification [41], and music categorization in terms of
emotion [42], among others. The possibility of multigenres classification is particularly
promising and probably closer to the human experience. Another interesting future work
is related to the synthesis of rhythms. Once the rhythmic networks are available, new
rhythms with similar characteristics according to the specific genre can be artificially
generated.
5. Acknowledgments
Debora C Correa is grateful to FAPESP (2009/50142-0) for financial support and
Luciano da F. Costa is grateful to CNPq (301303/06-1 and 573583/2008-0) and FAPESP
(05/00587-5) for financial support.
28
Appendices
Appendix A. Multivariate Statistical Methods
Appendix A.1. Principal Component Analysis
Principal Component Analysis is a second order unsupervised statistical technique.
By second order it is meant that all the necessary information is available directly
from the covariance matrix of the mixture data, so that there is no need to use the
complete probability distributions. This method uses the eigenvalues and eigenvectors
of the covariance matrix in order to transform the feature space, creating orthogonal
uncorrelated features. From a multivariate dataset, the principal aim of PCA is to
remove redundancy from the data, consequently reducing the dimensionality of the data.
Additional information about PCA and its relation to various interesting statistical
and geometrical properties can be found in the pattern recognition literature, e.g.
[43, 32, 7, 6].
Consider a vector x with n elements representing some features or measurements
of a sample. In the first step of PCA transform, this vector x is centered by subtracting
its mean, so that x ← x − E {x}. Next, x is linearly transformed to a different
vector y which contains m elements, m < n, removing the redundancy caused by the
correlations. This is achieved by using a rotated orthogonal coordinate system in such a
way that the elements in x are uncorrelated in the new coordinate system. At the same
time, PCA maximizes the variances of the projections of x on the new coordinate axes
(components). These variances of the components will differ in most applications. The
axes associated to small dispersions (given by the respectively associated eigenvalues)
can be discarded without losing too much information about the original data.
Appendix A.2. Linear discriminant Analysis
Linear discriminantAnalysis (LDA) can be considered a generalization of Fisher’s Linear
discriminant Function for the multivariate case [7, 32]. It is a supervised approach that
maximizes data separability, in terms of a simililarity criterion based on scatter matrices.
The basic idea is that objects belonging to the same class are as similar as possible and
objects belonging to distinct classes are as different as possible. In other words, LDA
looks for a new, projected, feature space where that maximizes interclass distance while
minimizing the intraclass distance. This result can be later used for linear classification,
and it is also possible to reduce dimensionality before the classification task. The scatter
matrix for each class indicates the dispersion of the features vectors within the class. The
intraclass scatter matrix is defined as the sum of the scatter matrices of all classes and
expresses the combined dispersion in each class. The interclass scatter matrix quantifies
how disperse the classes are, in terms of the position of their centroids.
It can be shown that the maximization criterion for class separability leads to
a generalized eigenvalue problem ([32, 7]). Therefore, it is possible to compute the
29
eigenvalues and eigenvectors of the matrix defined by(S−1
intra ∗ Sinter
), where Sintra is the
intraclass scatter matrix and Sinter is the interclass scatter matrix. The m eigenvectors
associated to the m largest eigenvalues of this matrix can be used to project the data.
However, the rank of(S−1
intra ∗ Sinter
)is limited to C−1, where C is the number of classes.
As a consequence, there are C−1 nonzero eigenvalues, that is, the number of new features
is conditioned to the number of classes, m ≤ C − 1. Another issue is that, for high
dimensional problems, when the number of available training samples is smaller than
the number of features, Sintra becomes singular, complicating the generalized eigenvalue
solution.
More information about the LDA is [31, 33, 7, 32].
30
Appendix B. Linear and Quadratic Discriminant Functions
If normal distribution over the data is assumed, it is possible to state that:
p (x|ωi) =1
(2π)d2 |Σ|
12
exp{−1
2(~x− ~µ)T Σ−1 (~x− ~µ)
}(B.1)
The components of the parameter vector for class j, ~θj = {~µj,Σj}, where ~µj and Σj
are the mean vector and the covariance matrix of class j, respectively, can be estimated
by maximum likelihood as follows:
~µj =1
N
N∑i=1
~xi (B.2)
~Σj =1
N
N∑i=1
(~xi − ~µi) (~xi − ~µi)T (B.3)
Within this context, classification can be achieved with discriminant functions, gi,
assigning an observed pattern vector ~xi to the class ωj with the maximum discriminant
function value. By using Bayes’s rule, not considering the constant terms, and using
the estimated parameters above, a decision rule can be defined as: assign an object ~xi
to class ωj if gj > gi for all i 6= j, where the discriminantfunction gi is calculated as:
gi (~x) = log (p (ωi))−1
2log
(∣∣∣Σi
∣∣∣)− 1
2
(~x− ~µj
)TΣ−1
i
(~x− ~µj
)(B.4)
Classifying an object or pattern ~x on the basis of the values of gi (~x), i = 1, ..., C (C
is the number of classes), with estimated parameters, defines a quadratic discriminant
classifier or quadratic Bayesian classifier or yet quadratic Gaussian classifier [32].
The prior probability, p (ωi), can be simply estimated by:
p (ωi) =ni∑j nj
(B.5)
where ni is the number of samples of class ωi.
In multivariate classification situations, with different covariance matrices,
problems may occur in the quadratic Bayesian classifier when any of the matrices
Σi is singular. This usually happens when there are not enough data to obtain
efficient estimative for the covariance matrices Σi, i = 1, 2, ..., C. An alternative to
minimize this problem consist of estimating one unique covariance matrix over all classes,
Σ = Σ1 = ... = ΣC . In this case, the discriminantfunction becomes linear in ~x and can
be simplified:
gi (~x) = log (p (ωi))−1
2~µ
T
i Σ−1~µi + ~xT Σ−1~µi (B.6)
where Σ is the covariance matrix, common to all classes. The classification rule
remains the same. This defines a linear discriminant classifier (also known as linear
Bayesian classifier or linear Gaussian classifier) [32].
31
Figure C1. A dendrogram for a simple situation involving eight objects, adaptedfrom [7].
Appendix C. Agglomerative Hierarchical Clustering
Agglomerative hierarchical clustering groups progressively the N objects into C classes
according to a defined parameter. The distance or similarity between the feature vectors
of the objects are usually taken as such parameter. In the first step, there is a partition
withN clusters, each cluster containing one object. The next step is a different partition,
with N − 1 clusters, the next a partition with N − 2 clusters, and so on. In the nth
step, all the objects form a unique cluster. This sequence groups objects that are more
similar to one another into subclasses before objects that are less similar. It is possible
to say that, in the kth step, C = N − k + 1.
To show how the objects are grouped, hierarchical clustering can be represented by a
corresponding tree, called dendrogram. Figure C1 illustrates a dendrogram representing
the results of hierarchical clustering for a problem with eight objects. The measure of
similarity among clusters can be observed in the vertical axis. The different number of
classes can be obtained by horizontally cutting the dendrogram at different values of
similarity or distance.
Hence, to perform hierarchical cluster analysis it is necessary to define three main
parameters. The first regards how to quantify the similarity between every pair of
objects in the data set, that is, how to calculate the distance between the objects.
Euclidean distance, which is frequently used, will be adopted in this work, but other
possible distances are cityblock, cheesboard, mahalanobis and so on. The second
parameter is the linkage method, which establishes how to measure the distance between
two sets. The linkage method can be used to link pairs of objects that are similar and
then to form the hierarchical cluster tree. There are many possibilities for doing so,
some of the most popular are: single linkage, complete linkage, group linkage, centroid
linkage, mean linkage and ward’s linkage ([44, 45, 46, 6]). Ward’s linkage uses the
intraclass dispersion as a clustering criterion. Pairs of objects are merged in such a
way to guarantee the smallest increase in the intraclass dispersion. This clustering
approach has been sometimes identified as corresponding to the best hierarchical method
32
[47, 48, 49] and will be used in this work. Actually, it is particularly interesting to analyse
the intraclass dispersion in an unsupervised classification procedure in order to identify
common and different characteristics when compared to the supervised classification.
The third parameter concerns the number of desired clusters, an issue which is directly
related to where to cut the dendrogram into clusters, as illustrated by C in Figure C1.
33
Table D1. Classification performance according to kappaKappa Classification performance
k ≤ 0 Poor0 < k ≤ 0.2 Slight0.2 < k ≤ 0.4 Fair0.4 < k ≤ 0.6 Moderate0.6 < k ≤ 0.8 Substantial0.8 < k ≤ 1.0 Almost Perfect
Appendix D. The Kappa Coefficient
The kappa coefficient was first proposed by Cohen [35]. In the context of supervised
classification, this coefficient determines the degree of agreement a posteriori. This
means that it quantifies the agreement between objects previously known (ground truth)
and the result obtained by the classifier. The better the classification accuracy, the
higher the degree of concordance, and, consequently, the higher the value of kappa. The
kappa coefficient is computed from the confusion matrix as follows [36]:
k =N∑C
i=1 cii −∑C
i=1 xi+x+i
N2 −∑Ci=1 xi+x+i
(D.1)
where xi+ is the sum of elements from line i, x+i is the sum of elements from column
i, C is the number of classes (confusion matrix is C x C), and N is the total number of
objects. The kappa variance can be calculated as:
σ2k =
1
N
[θ1 (1− θ1)
(1− θ2)2 +
2 (1− θ1) (2θ1θ2 − θ3)
(1− θ2)3 +
(1− θ1)2 (θ4 − 4θ2
2)
(1− θ2)4
](D.2)
where
θ1 =1
N
C∑i=1
xii
θ2 =1
N2
C∑i=1
xi+x+i
θ3 =1
N2
C∑i=1
xii (xi+ + x+i)
θ4 =1
N3
C∑i=1
C∑j=1
xij (xj+ + x+i)2
(D.3)
This statistics indicates that, when k ≤ 0 there is not any agreement, and when
k = 1 the agreement is total. Some authors suggest interpretations according to the
value obtained by the coefficient kappa. Table D1 shows one possible interpretation,
proposed by [50].
34
References
[1] J. C. LENA and R. A. PETERSON. Classification as culture: Types and trajectories of musicgenres. American Sociological Review., 73:697–718, 2008.
[2] F. HOLT. Genre In Popular Music. University of Chicago Press, 2007.[3] N. SCARINGELLA, G. ZOIA, and D. MLYNEK. Automatic genre classification of music content
- a survey. IEEE Signal Processing Magazine, pages 133–141, 2006.[4] J.J. AUCOUNTURIER and PACHET F. Representing musical genre: A state of the art. Journal
of New Music Research, pages 83–93, 2005.[5] F. PACHET and D. CAZALY. A taxonomy of musical genres. In Proc. Content-Based Multimedia
Information Acess (RIAO), 2000.[6] L. da F. COSTA and R. M. CESAR JR. Shape Analysis and Classification. Theory and Practice.
CRC Press, 2001.[7] R. O. DUDA, P. E. HART, and D. G. STORK. Pattern Classification. John Wiley & Sons, Inc.,
2001.[8] Z. CATALTEPE, Y. YASIAN, and A. SONMEZ. Music genre classification using midi and audio
features. EURASIP Journal on Advances of Signal Processing, pages 1–8, 2007.[9] T. LI, M. OGIHARA, B. SHAO, and D. WANG. Theory and Novel Applications of Machine
Learning, chapter Machine Learning Approaches for Music Information Retrieval, pages 259–278. I-Tech, Vienna, Austria., 2009.
[10] M. M. MOSTAFA and N. BILLOR. Recognition of western style musical genres using machinelearning techniques. Expert Systems with Applications, 36:11378–11389, 2009.
[11] L. WANG, S. HUANG, S. WANG, J. LIANG, and B. XU. Music genre classification based onmultiple classifier fusion. Fourth International Conference on Natural Computation, pages 580–583, 2008.
[12] Y. SONG and C. ZHANG. Content-based information fusion for semi-supervised music genreclassification. IEEE Transactions on Multimedia, 10:145–152, 2008.
[13] I. PANAGAKIS, E. BENETOS, and C. KOTROPOULOS. Music genre classification: Amultilinear approach. Proceedings of the Ninth International Conference on Music InformationRetrieval., pages 583–588, 2008.
[14] J. HONG, H. DENG, and Q. YAN. Tag-based artist similarity and genre classification. Proceedingsof the IEEE International Symposium on Knowledge Acquisition and Modeling., pages 628–631,2008.
[15] A. HOLZAPFEL and Y. STYLIANOU. A statistical approach to musical genre classificationusing non-negative matrix factorization. IEEE International Conference on Acoustics, Speechand Signal Processing, 2:II–693–II–696, 2007.
[16] C. N. SILLA, A. A. KAESTNER, and A. L. KOERICH. Automatic music genre classificationusing ensemble of classifiers. Proceedings of the IEEE International Conference on Systems,Man and Cybernetics., pages 1687–1692, 2007.
[17] N. SCARINGELLA and G. ZOIA. On the modeling of time information for automatic genrerecognition systems in audio signals. Proceedings of the 6th International Symposium on MusicInformation Retrieval, pages 666–671, 2005.
[18] X. SHAO, C. XU, and M. KANKANHALLI. Unsupervised classification of musical genre usinghidden markov model. Proceedings of the International Conference on Multimedia Explore(ICME), pages 2023–2026, 2004.
[19] J. J. BURRED and A. LERCH. A hierarchical approach to automatic musical genre classification.Proceedings of the International Conference on Digital Audio Effects, pages 8–11, 2003.
[20] Md. AKHTARUZZAMAN. Representation of musical rhythm and its classification system basedon mathematical and geometrical analysis. Proceedings of the International Conference onComputer and Communication Engineering, pages 466–471, 2008.
[21] I. KARYDIS. Symbolic music genre classification based on note pitch and duration. Proceedings
35
of the 10th East European Conference on Advances in Databases and Information Systems.,pages 329–338, 2006.
[22] F. GOUYON and S. DIXON. A review of automatic rhythm description system. Computer MusicJournal, 29:34–54, 2005.
[23] F. GOUYON, S. DIXON, E. PAMPALK, and G. WIDMER. Evaluating rhythmic descriptors formusical genre classification. Proceeding of the 25th International AES Conference, 2004.
[24] R. ALBERT and A. L. BARABSI. Statistical mechanics of complex networks. Reviews of ModernPhysics, 74:47–97, 2002.
[25] M. E. J. NEWMAN. The structure and function of complex networks. SIAM Review, 45:167–256,2003.
[26] T. L. BOOTH. Sequential Machines and Automata Theory. John Wiley & Sons Inc., 1967.[27] M. ROHRMEIER. Towards Modelling Harmonic Movement in Music: Analysing Properties and
Dynamic Aspects of Pc Set Sequences in Bach’s Chorales. PhD thesis, Cambridge University,2006.
[28] L. da F. COSTA, F. A. RODRIGUES, G. TRAVIESO, and P. R. VILLAS BOAS. Characterizationof complex networks: A survey of measurements. Advances in Physics., 56(1):167–242, 2007.
[29] E. R. MIRANDA. Composing Music With Computers. Focal Press, 2001.[30] T. EEROLA and P.I. TOIVIAINEN. MIDI Toolbox: MATLAB Tools for Music Research.
University of Jyvaskyla, Jyvaskyla, Finland, 2004.[31] P. A. DEVIJVER and J. KITTLER. Pattern Recognition: A Statistical Approach. Prentice Hall
International, 1981.[32] A. R. WEBB. Statistical Pattern Recognition. John Wiley & Sons Ltd, 2002.[33] C. W. THERRIEN. Decision Estimation and Classification: An Introduction to Pattern
Recognition and Related Topics. John Wiley & Sons, 1989.[34] S. THEODORIDIS and K. KOUTROUMBAS. Pattern Recognition. Elsevier (USA), 2006.[35] J. COHEN. A coefficient of agreement for nominal scales. Educational and Psychological
Measurement, 20, n.1:37–46, 1960.[36] R. G CONGALTON. A review of assessing the accuracy of classifications of remotely sensed data.
Remote Sensing of Enviroment, 37:35–46, 1991.[37] K. FUKUNAGA. Introduction to Statistical Pattern Recoginition. Academic Press, 1990.[38] G. TSOUMAKAS and I. KATAKIS. Multi-label classification: An overview. International Journal
of Data Warehousing and Mining., 3(3):1–13, 2007.[39] I. KATAKIS, G. TSOUMAKAS, and I. VLAHAVAS. Multilabel text classification for automated
tag suggestion. Proceedings of the ECML/PKDD 2008 Discovery Challenge, 2008.[40] M. R. BOUTELL, J. LUO, X. SHEN, and C. M. BRWON. Learning multi-labelscene
classiffication. Pattern Recognition, 37(9):1757–1771, 2004.[41] S. DIPLARIS, G. TSOUMAKAS, P. MITKAS, and I. VLAHAVAS. Protein classification with
multiple algorithms. Proc. 10th Panhellenic Conference on Informatics (PCI 2005), pages 448–456, 2005.
[42] K. TROHIDIS, G. TSOUMAKAS, G. KALLIRIS, and I. VLAHAVAS. Multilabel classificationof music into emotions. Proc. 2008 International Conference on Music Information Retrieval(ISMIR 2008), pages 325–330, 2008.
[43] A. HYVARINEN, J. KARHUNEN, and E. OJA. Independent Component Analysis. John Wiley& Son, Inc., 2001.
[44] A. K. JAIN and R. C. DUBES. Algorithms for Clustering Data. Prentice Hall, 1988.[45] M. R. ANDERBERG. Cluster Analysis for Applications. Academic Press, 1973.[46] H. C. ROMESBURG. Cluster Analysis for Resarches. LULU Press, 1990.[47] F. K. KUIPER and L. A. FISHER. Monte carlo comparison of six clustering procedures.
Biometrics, 31:777–783, 1975.[48] R. K. BLASHFILED. Mixture model tests of cluster analysis: Accuracy of four agglomerative